In recent years, sentiment classification has succeeded greatly in high-resource languages such as English and Chinese. However, most languages do not benefit from such rich annotated data, particularly low-resource languages. To address the sentiment classification problem in low-resource languages lacking sufficient annotated data, we augment the data in two steps by transferring knowledge learned from labeled data in a high-resource source language to low-resource languages. First, we propose a sentiment word-based control strategy that enables pre-trained language models to generate diverse and sentiment-rich augmented data. Second, we construct code-switched text using a bilingual sentiment lexicon, encouraging the model to align representations from the source and multiple target languages by blending their contextual information. We evaluate the effectiveness of our approach on two self-constructed Chinese minority language datasets and achieve the best performance compared to the baselines.